DeepSeek is expanding its software to run AI models more efficiently on Huawei chips. To this end, the Chinese AI developer is making programming tools for the Ascend platform available as open source. In doing so, DeepSeek and Huawei are targeting not only alternative AI hardware but also one of Nvidia’s key strengths: the software ecosystem surrounding its GPUs.
Nvidia owes its dominant position in AI to more than just powerful processors. CUDA, the programming platform that developers use to build and optimize software for Nvidia GPUs, poses a significant barrier to competitors. After all, anyone wanting to use different hardware must also have access to suitable development software and libraries.
DeepSeek is attempting to lower that barrier for Huawei’s Ascend processors. According to Reuters, the companies are making available, among other things, computing and communication libraries optimized for the Ascend platform. Huawei has supported DeepSeek in its development.
TileLang, an open-source programming language for AI chips, plays a central role. It lets developers specify which computations to perform at a higher level of abstraction, while the software translates them to the underlying hardware. DeepSeek claims that this provides a simpler programming model than CUDA.
Multiple vendors
According to the South China Morning Post, this involves a broader suite of core software that has made DeepSeek compatible with Ascend. In doing so, the company is bringing technology it previously developed for Nvidia hardware to Huawei’s platform. The ultimate goal is to make developers less dependent on a single supplier of AI accelerators.
The collaboration is not limited to programming tools. DeepSeek and Huawei have also worked on a supernode featuring 128 Ascend 950 processors. Reuters says both the distribution of computations and the communication between the chips have been optimized.
Two weeks ago, Huawei unveiled a new generation of Ascend processors and systems that combine large numbers of these chips. The company had previously made it clear that it intends to further intensify competition with Nvidia using its own HBM memory and new SuperPod technology. Huawei expects such systems to be used more frequently for training AI models starting next year.
For Huawei, the combination of hardware and software matters most. Simply developing a processor that can compete with Nvidia is not enough if developers then remain dependent on software written specifically for CUDA. Through their collaboration on TileLang and the associated libraries, Huawei and DeepSeek are also working to reduce that dependency.